Title: Zero-Knowledge Cloud Analytics
Authors: Zeying Zhu, Clarence Lam, Alexander Frolov, Ian Miers, Zhaoxing Liu (University of Maryland)
Introduction
This paper studies how cloud analytics results can be publicly verifiable without exposing raw logs. The problem is important because many cloud services publish metrics such as video QoE, network performance, or compliance statistics, but those results are often self-asserted. External auditors usually cannot access the raw data, and existing solutions are not fully satisfactory: TEEs still rely on trusted hardware, while end-user measurement only gives partial visibility.
Key idea and contribution
The authors propose zk-Analytics, which splits the analytics pipeline into three stages: online append-only log commitment at the data source, offline distributed aggregation, and offline query verification. The core idea is to first commit to logs using lightweight hash chains, then run aggregation and query logic inside zk-SNARKs / a zkVM, so that any verifier can check that the published result was computed correctly from the committed logs.
The system is designed as an end-to-end pipeline rather than just a proof circuit. It supports multiple analytics styles, including raw-log queries, hash tables, histograms, and Count-Min Sketches. The main engineering trick is to keep the online path light and push heavy computation into the offline proving stage, which makes the system scalable enough to be practical for batch analytics and audit-style verification.
Evaluation
The authors implemented the system in Rust using RISC Zero zkVM, with about 13K lines of code, and evaluated it on Google cluster traces, CAIDA traffic, and a vehicle emissions dataset. The results show that online log commitment is very cheap and verification is lightweight, usually under 100 ms. The main cost is proof generation: aggregation proofs can take around 90.8 minutes on Google cluster data and 227.5 minutes on CAIDA data. Proof sizes remain relatively compact, with aggregation proofs around 1-2 MB and query proofs around 222-225 KB.
This result is significant because it shows that publicly verifiable cloud analytics is feasible on real workloads, not just a theoretical idea. In other words, the paper moves the trust model from “you have to believe the provider” to “you can verify the provider’s output,” which is valuable for auditing, compliance, and cross-organization reporting.
Q: Can the system support data providers that intentionally obscure or hide statistics, such as traffic volume, number of users, or failure counts, in order to protect sensitive information?
A: This scenario involves a different threat model. The current work assumes that the data sources are trusted. To protect the statistics from potentially untrusted data sources, it may be necessary to deploy Trusted Execution Environments (TEEs) at every data factory.
Personal thoughts
I think the strongest part of this paper is that it defines the problem clearly and gives a very coherent system design. The authors do not try to solve everything at once; instead, they focus on making analytics results verifiable, which is a realistic and useful goal. I also like the clean separation between online commitment and offline proving, since that matches how many log-driven analytics systems actually operate.
That said, the cost of proof generation is still very high, especially for larger workloads, so this looks more like an audit or post-hoc verification system than a real-time interactive one. Another limitation is that the system still reveals some structural metadata, such as batch boundaries and source counts. For future work, I would be interested in seeing whether proof generation can be made much cheaper, and whether metadata leakage could be reduced further, possibly by combining zero knowledge with differential privacy in a more integrated way.